Glossary

Correlates inversely with reliance on free AI tools for health information. The Commonwealth Fund's 2025 Health Care Affordability Survey found that 43 percent of working-age adults skipped or delayed care due to cost, with uninsured and low-income populations most affected.

AI systems that can execute multi-step tasks autonomously, not just answer questions. They range from open-source research tools like Hermies to full writing-and-publishing pipelines. The trade-off is complexity: agents require configuration, clear instructions, and a human who understands what the agent is doing well enough to catch when it goes wrong. They are not plug-and-play.

Extends the GPL v3 with Section 13, which requires that anyone running AGPL-licensed software as a network service (SaaS, web application, API) must offer the source code to users interacting with it remotely. This closes the "SaaS loophole" – the gap that lets cloud providers run GPL code as a service without triggering the share-alike requirement because no binary is technically "distributed." gnu.org/licenses/agpl-3.0.html.

Independent benchmark ranking AI models across standardised evaluations. Command A ranked #17 out of 38 open-weight non-reasoning models. Intelligence score: 13 (median: 13 – exactly average). Full benchmark data: artificialanalysis.ai/models/command-a. Compare directly with free alternatives: DeepSeek-R1 (higher intelligence score, $0 cost), Llama 4 Maverick (higher score, open-weight, commercially licensed), Qwen3 (comparable score, open-weight, toggleable reasoning). Full leaderboard: artificialanalysis.ai/leaderboards.

A nonpartisan public policy think tank founded in 1916 in Washington, D.C. Its 2026 report on AI in K-12 education concluded that the risks of generative AI in education overshadow current benefits, particularly regarding student-teacher relationships and student safety.

A $2 billion initiative announced in Budget 2024 to build domestic AI infrastructure. Three pillars: AI Compute Challenge ($700M for private-sector projects), AI Compute Access Fund ($300M for startup access), and Public Infrastructure ($705M for government-owned supercomputing). ised-isde.canada.ca.

U.S. law signed March 2018, 18 U.S.C. § 2713. Allows American law enforcement to compel U.S.-headquartered companies to produce data within their "possession, custody, or control" regardless of where that data is physically stored. The determining factor is the company's connection to U.S. jurisdiction, not the location of the servers.

All three licences (MIT, GPL v3, AGPL v3) are OSI-approved open-source licences. All permit commercial use, modification, and distribution. The critical difference is the share-alike requirement: MIT has none, GPL requires it for distributed works, AGPL requires it for distributed works AND network services.

A federal law enacted in 1998 (15 U.S.C. §§ 6501–6506) that requires operators of websites and online services directed to children under 13 to obtain verifiable parental consent before collecting, using, or disclosing personal information from children. The FTC enforces COPPA through the COPPA Rule (16 CFR Part 312). ftc.gov/coppa.

Student Data Privacy Toolkit covering vendor vetting, edtech inventories, and data governance. cosn.org.

The Korean College Scholastic Ability Test (Suneung), a national standardised exam taken annually by approximately 500,000 students for university admission. AI model scores on CSAT are reported in the DCO Digital Economy Trends 2026 report. det.dco.org

The practice of using your own product internally before releasing it to customers. Companies release new features to employees first to catch bugs and usability problems before real users encounter them. The term has been in use in the technology industry since at least the 1980s.

The phenomenon where AI tools inflate users' confidence in their own competence beyond what their actual performance warrants. Research published on arXiv found that AI-assisted groups reported significantly higher self-assessment scores regardless of objective results. arxiv.org

Doctorow's term for the pattern by which platforms first attract users with quality, then extract value from users to attract business customers, then extract value from both to maximise shareholder returns. Each stage degrades the experience of the previous group.

A benchmark for evaluating emotional intelligence in large language models, introduced in 2023 by researchers who designed sixty English-language dialogue snippets portraying conflict and nuanced social interaction. Models predict the intensity of emotional states in characters. The benchmark's strong correlation with MMLU (r=0.97) suggests emotional performance in LLMs is a dimension of general capability, not a separate faculty. Paech, S. J. "EQ-Bench: An Emotional Intelligence Benchmark for Large Language Models." arXiv, 2023. arxiv.org/abs/2312.06281

A federal law enacted in 1974 (20 U.S.C. § 1232g) that protects the privacy of student education records. Schools must accommodate inspection requests within 45 days. The Department of Education has stated that "schools cannot require parents or students to waive their FERPA rights through ed tech company's terms of service." studentprivacy.ed.gov.

The EU General Data Protection Regulation (2018) requires explicit, informed consent before processing children's personal data. Article 8 sets a default age of 16, with member states permitted to lower it to 13. gdpr.eu/children.

A copyleft licence consisting of approximately 5,600 words. It permits the same uses as MIT but additionally requires that any distributed derivative works be licensed under the same GPL terms, ensuring that modifications remain available to the community. Published by the Free Software Foundation in 2007. gnu.org/licenses/gpl-3.0.html.

A crowdsourced benchmark platform using 6M+ user votes to compute Elo ratings for language models. Cohere's Command R+ ranked #37 as of September 2025 — below free models from DeepSeek, Meta (Llama), and Mistral. Live leaderboard: lmarena.ai/leaderboard. For comparison: Claude Opus 4.6 ranks #1-2, GPT-5.2 ranks #1-2, Gemini 3.1 Pro ranks #3. Cohere is not in the conversation.

A permissive open-source licence consisting of approximately 170 words. It permits use, copying, modification, merging, publishing, distributing, sublicensing, and selling copies of the software, with the sole condition that the copyright notice and permission notice are included in all copies. It imposes no obligation to share modifications or provide source code.

Permissive licences (MIT, Apache, BSD) place minimal restrictions on use, including allowing incorporation into proprietary software. Copyleft licences (GPL, AGPL, MPL) require that derivative works be shared under the same or compatible licence, preventing proprietary capture of open-source improvements.

Formalized by Shigeo Shingo in 1961 as part of the Toyota Production System. Shingo distinguished between mistakes (inevitable human errors) and defects (mistakes that reach the customer). The goal of poka-yoke is to design processes so that mistakes are detected and corrected immediately, eliminating defects at the source. Originally called baka-yoke ("fool-proofing"), the name was changed in 1963 after a worker at Arakawa Body Co. objected to the term.

The National Data Privacy Agreement (NDPA), Version 2, developed by 28 state alliances. privacy.a4l.org.

Enacted 1914: prohibits a person from serving simultaneously as a director or officer of two competing corporations when each company's combined capital, surplus, and undivided profits exceeds the jurisdictional threshold ($48,559,000 in 2024, indexed annually).

The practice of moving testing activities earlier in the software development lifecycle, embedding quality checks into development rather than performing them at the end. The term originated in the testing community in the early 2010s and was widely adopted by 2015. The philosophy is sound; the problem was that what shifted left was automated testing, while systemic quality assurance fell out of scope.

An open-source continuous delivery platform originally built by Netflix and later adopted by Google. It automates the process of releasing software to production environments, handling deployment strategies, rollbacks, and multi-cloud orchestration.

First named in a series of NTIA reports beginning in 1995 ("Falling Through the Net"), which documented disparities in computer and internet access by income, race, geography, and education. The term entered mainstream policy vocabulary by the late 1990s and shaped federal programmes including E-Rate (1996), which subsidised internet access for schools and libraries.

The term coined by psychologist Betsy Sparrow and colleagues at Columbia University in a 2011 study published in Science. The research demonstrated that when people expect to have future access to information (e.g., via a search engine), they show lower rates of recall for the information itself but enhanced recall for where to find it. The study established that search engines function as an external memory system, changing not what we know but how we know it. Sparrow, B., Liu, J., & Wegner, D. M. (2011). "Google Effects on Memory: Cognitive Consequences of Having Information at Our Fingertips." Science, 333(6043), 776-778. doi.org/10.1126/science.1207745

Enacted 2021, enforced by the Information Commissioner's Office (ICO). Requires that digital services likely to be accessed by children provide privacy protections by default. Applies to any service accessible by children in the UK, regardless of where the company is headquartered. ico.org.uk.

A human-readable data format used for configuration files. It uses indentation and simple punctuation rather than the brackets and braces of formats like JSON. In the context of AI agent frameworks, YAML files define the rules and workflows that control how agents operate, what steps they follow, and what constraints they must respect.